Define retry policy when generation request timeouts

[OrGivati]

We are using gemini using ai-sdk(@ai-sdk/google-vertex:4.0.15) and also some of the OpenAI models.
From time to time, some generation request take too long - +300 second(see image).

My questions:

1. can I control the timeout of the generation request?
2. can I define retry policy for the agent to apply in case generation fails on timeout? or other exceptions?

if we will be able to set to the run options(SharedRunOptions) retry policy on generation request based on the errors type thrown by the provider, it can help improve the reliability of the agent run.
Both by providing custom timeout I can fail request early and with the retry policy I can recover in case of failures.

Debug information

• Agents SDK version: (e.g. v0.4.0)
• Runtime environment (e.g. Node.js 22.19.0)

[seratch]

You can enforce timeouts via SharedRunOptions.signal. The ai-sdk adapter forwards it to its abortSignal. With node 22, you can do:

const result = await run(agent, input, {
  signal: AbortSignal.timeout(30_000),
});

If you need broader support, use AbortController + setTimeout and pass its signal.

There is no built‑in run‑level retry policy in agents-core package right now. It seems ai-sdk supports maxRetries in call options, and the adapter already spreads modelSettings.providerData into ai-sdk call options — so you can set:

const agent = new Agent({
  // ...
  modelSettings: {
    providerData: { maxRetries: 2 },
  },
});

For future ideas of the customizable retry policy, we may consider adding new options for it, but I'd like to learn what's the most useful options for many developers, an how ai-sdk could support something similar.

[OrGivati]

Thanks!

I will use the provider data to define the timeout and retries.

I want to clarify, the retry policy and timeouts I want to define are not for the entire agent run, but specific request to the model during the run which take abnormally longer than usual.
Since the agent can run for minutes - losing the entire run because 1 request that failed on timeout / any other known exception is a shame.
Define retry policy in the SDK level can help a lot to improve the reliability of the application. Although, I get it might be depend on the model provider and its specific exceptions.
The SDK might help ensure that I won't apply retries on every generation request and might just backoff entirely from the run.

[seratch]

Thanks for the context. For OpenAI models, adding timeout and other custom settings per model call could be relatively quick. Adding new options to ModelSettings or elsewhere should work well. As you mentioned, supporting the exactly same behavior for all available models could be a challenge. I haven't checked the feasibility yet. I will continue doing research on this topic.